The Long-Run Impacts of a Universal Child Care Program
By MICHAEL BAKER, JONATHAN GRUBER, AND KEVIN MILLIGAN*
ACCEPTED, AEJ: ECONOMIC POLICY
THIS VERSION: NOVEMBER 8, 2018
Past research documents the persistence of positive impacts of
early-life interventions on non cognitive skills. We test the symmetry
of this finding by studying the persistence of a sizeable negative
shock to non cognitive outcomes arising with the introduction of
universal child care in Quebec. We find that the negative effects on
non cognitive outcomes persisted to school ages, and also that
cohorts with increased child care access had worse health, lower
life satisfaction, and higher crime rates later in life. Our results
reinforce previous evidence of the central role of the early
childhood environment for long-run success. (JEL I1, J13, K42)
* Baker: Department of Economics, University of Toronto, 150 St. George St., Toronto ON M5S 3G7, and NBER (e-mail:
[email protected]); Gruber: Department of Economics, Massachusetts Institute of Technology, 50 Memorial Drive,
Cambridge, MA 02142, and NBER (e-mail: [email protected]); Milligan: Vancouver School of Economics, University of
British Columbia, 6000 Iona Drive, Vancouver BC V6T 1L4, and NBER (e-mail: [email protected]). We thank
Hélène Durocher, Pablo Gutierrez Cubillos, and Timea Molnar for outstanding research assistance and seminar participants
at McMaster, the Norwegian Research Council workshop “Interventions during childhood and subsequent child
development”, Simon Fraser, University of Victoria, and the World Bank. Much of the analysis for this paper was
conducted at the British Columbia Interuniversity Research Data Centre, which is part of the Canadian Research Data
Centre Network (CRDCN). The services and activities provided by the CRDCN are made possible by the financial or in-
kind support of the SSHRC, the CIHR, the CFI, Statistics Canada and participating universities whose support is gratefully
acknowledged. The views expressed in this paper do not necessarily represent the CRDCN’s or that of its partners. Baker
gratefully acknowledges the research support of SSHRC (Grant, #410-2011-0724) and a Canada Research Chair at the
University of Toronto.
1
An enduring question about child development is the persistence of early
childhood interventions. The debate on Head Start in the United States, for
example, has focused in part on whether positive effects of the program found at
younger ages fade through time. For example, Gibbs et al. (2013), Bitler et al.
(2014), and Kline and Walters (2016) use data from a randomized study of Head
Start and generally find the initial cognitive impact fades by first grade. Kline and
Walters (2016) argue, however, that even that initial impact can have long-run
impact on earnings. Using a regression discontinuity approach, Carneiro and
Ginja (2014) found sizeable and persistent long-run benefits from Head Start
program participation. With a similar focus on persistence, the re-examination of
evidence on the Perry Preschool Project by Heckman et al. (2013) distinguished
between positive cognitive effects which faded and positive non cognitive effects
which persisted. These latter long-run impacts led to improved economic
outcomes and lower incidence of criminal behavior, such that Heckman et al.
(2010) find the measured annualized rate of return to the Perry Preschool
investment is 6-10%.
These findings leave untouched an important question of symmetry: are there
equally persistent and important negative long-run impacts of interventions that
had an initial negative impact? This question is important because an affirmative
answer would buttress the general case for the importance of the early childhood
environment. When a supportive developmental environment is present, children
may benefit in the long run. This case is made stronger with evidence that a
deficient early environment has a long-run detrimental impact. An example of this
latter type of evidence is Bertrand and Pan (2013) who found that childhood non
cognitive deficits contribute to the gender difference in teenage disruptive
behavior. Also, investigations into the symmetry of early-life health and human
2
capital events have also reinforced the importance of looking at both positive and
negative shocks in related early-life contexts.1
In this paper, we develop a causal estimate of the long-run impact of a child
care intervention on long-run later-life outcomes. To do this we study the largest
experiment with universal child care in North America in recent years: an
introduction of very low-cost child care for children aged 0-4 in Quebec in 1997.
In an earlier paper (Baker, Gruber, and Milligan 2008; henceforth BGM) we
documented large increases in maternal labor supply and in the placement of
children in child care in Quebec relative to the rest of Canada, where child care
services remained unchanged. (See also Lefebvre and Merrigan 2008 and
Lefebvre et al. 2009.) At the same time, there was a large, significant, negative
shock to the preschool, non cognitive development and health of children exposed
to the new program, with little measured impact on cognitive skills. Subsequent
research (Kottelenberg and Lehrer 2013) has confirmed that the negative
contemporaneous impact of the program on young children’s non cognitive
development has persisted as the program has matured.
Our analysis extends the study of the impact of this universal child care
program to children at older ages, looking for any persistence of the
contemporaneous impacts in four spheres: non cognitive development, cognitive
development, health, and crime. We begin our investigation by replicating
previous results showing that exposure to the Quebec program increased use of
child care among children age 0-4 and led to lower non cognitive outcomes at
those ages. We then proceed to our examination of persistence. We provide new
evidence for children aged 5-9, showing the program’s negative effect on non
cognitive skills do not appear to have faded by those ages—and in some cases are
1 For example, studies of early-life or in-utero infections are reviewed in Almond and Currie (2011), while negative
nutrition shocks from fasting are reviewed in Almond, Currie, and Duque (forthcoming). For human capital, Almond, Edlund, and Pålme (2009) find negative long-run human capital implications from an early-life negative shock to child
development caused by the Chernobyl nuclear accident.
3
even stronger. In this way, our results are a mirror-image of the Perry Preschool
and Head Start evidence. We also extend previous evidence of the heterogeneous
contemporaneous impact of the program. At older ages the primary, negative
impact of the program in on children who are already struggling.
We next explore the impacts of this child care intervention on outcomes in the
preteen and teenage years. The oldest children who were eligible for the Quebec
program are currently in their early 20s, and those who received full treatment are
now in their teens. We therefore focus, as the data allow, on their cognitive
achievement and health at these ages. We ask whether any early negative impacts
of the Program on these children’s cognitive development fade out, as has been
found in evaluations of other early-years programs? The program also led to
earlier exposure to infections and diseases and increases in early childhood
anxiety and aggression, and we investigate whether these provide any immunity
against maladies or have any consequences for life satisfaction and mental health,
respectively, at older ages.
Finally, we also look for any longer-term impacts of the early childhood deficits
in non cognitive development on criminal activity. This line of inquiry is guided
by the evidence from programs such as Perry Preschool which have shown how
positive, early childhood, impacts on non cognitive development result in more
positive interactions with various social institutions including the criminal justice
system.
Our results find no consistent evidence of a lasting impact of the Quebec
program on cognitive test scores; the available data give opposing answers for
math scores and show little effect on English or science scores. We do, however,
find a significant decline in self-reported health and in life satisfaction among
teens. Most strikingly, we find a sharp and contemporaneous increase in criminal
behavior among the cohorts exposed to the Quebec program, relative to their
peers in other provinces. We illustrate graphically a monotonic increase in crime
4
rates among cohorts with their exposure to the child care program, and we show
in regression analysis that exposure led to a significant rise in overall crime rates.
We also report that these effects are primarily for boys, who also see the largest
deterioration in non cognitive skills.
By showing that early shocks to childhood development can have persistent
effects, our results reinforce previous research emphasizing the importance of
early development for later-life outcomes and also provide an important input to
the current debate over child care policy. The rapid growth in female labor force
participation has led policy makers around the world to consider increased public
entitlement to child care for two-worker families. As one example, New York
City is implementing universal pre-kindergarten for three and four-year olds.2 The
evidence presented in this paper suggests that measurement of the near-term
impact of these policies can serve as a key indicator of the likely long-run success
or failure of similar programs.
Our paper proceeds as follows. Part I provides a summary of the extant
literature on child care and child outcomes. Part II discusses the Quebec reform.
Part III then introduces the wide variety of data sources that we will use for the
analysis and discusses our empirical strategy. Part IV updates the results on the
contemporaneous impact of the program, and Part V presents our results on the
persistence of the impact. Part VI concludes.
I. Background
There is now an enormous literature on the impacts of child care and preschool
on the outcomes of young children, and a smaller literature that examines any
longer-run impacts as the children age. Two important distinctions have emerged
interpreting the evidence. First is whether the child care intervention being studied
2 See http://www1.nyc.gov/office-of-the-mayor/news/258-17/mayor-de-blasio-3-k-all#/0
5
was targeted at children in more disadvantaged families or was universal and
targeted at all families. Second is the impact on cognitive or non cognitive
outcomes, both in the short and long run. We review this literature to place our
results in context, with an emphasis on recent research.3
A recent view of effects of previous child care exposure on outcomes in
adolescence suggest that more hours in child care in general does not affect test
scores, but has a negative effect on non cognitive outcomes, such as impulsivity
and risk-taking (Vandell et al., 2010). That study, typical of many in the literature,
relies on parental choice of child care mode, raising the question of whether any
estimated impacts of child care mode are causal or due to selection by parent type.
Similar problems affect the interpretation of the large existing literature in
economics on maternal work and child outcomes.
A growing body of evidence comes from the use of experimental and quasi-
experimental methods to examine the impacts of child care. Perhaps best known
are programs targeted toward at-risk children; for example, the experimental
variation embedded in the evaluations of the Abecedarian and Perry Preschool
interventions. These randomized trials from the 1960s and 1970s have shown that
high quality pre-school targeted to low-income children has substantial positive
effects. For example, Heckman et al. (2010) estimate a statistically significant
annual return of between 7 and 10 percent for the Perry Preschool intervention.
Carneiro and Heckman (2003) summarize the evidence from these programs as
improving motivation and social skills, while reducing crime and related
3 See a review of the literature up to 2008 in BGM, and in Baker (2011), Cascio (2015), and
Almond, Currie, and Duque (forthcoming). After completing our paper, we became aware that
Haeck et al. (2018) have published a paper investigating long term impacts of the Quebec program
on health, behavior, and motor-social development. They report that the contemporaneous
program effects on non cognitive skills persisted as the program matured and that effects for
“emotional disorder” and anxiety persist at school ages, although with smaller magnitudes.
6
behavior. Heckman et al. (2013) also argue that the non cognitive improvements
were pivotal to the long-run impact.
Unlike the experimental evaluations of model programs, our paper focuses on a
universal program that services a more economically and socially diverse group
of children. In contrast to the literature on programs targeting at-risk children, the
evidence on broader programs is mixed (see Baker 2011 and Cascio 2015 for
recent overviews). As examples we summarize the findings of evaluations of
programs in Denmark, Norway, Spain and Germany.
Exploiting variation in access to center-based preschool (versus a family-based
alternative) in Denmark, Datta Gupta and Simonsen (2010) report little effect on
non cognitive outcomes at age 7, and a negative impact of family-based child care
for boys of parents with low education. Black et al. (2014) utilize a discontinuity
in the price of child care in Norway, reporting that while neither child care
utilization nor parental labor supply is sensitive to price, they observe a positive
impact on children’s junior high school outcomes, presumably through a
disposable income effect. Havnes and Mogstad (2011) explore an expansion of
the Norwegian system, reporting positive impacts. The public system led to
higher educational attainment (primarily for children of low education mothers)
and earnings (mostly for girls) at ages 30–40. In a related paper Havnes and
Mogstad (2014) find that the earnings gains are primarily for children of low-
income parents and that children of upper-class parents experience an earnings
loss. Felfe et al. (2015) exploit variation across states in the expansion of the
Spanish child care system, finding improvements in reading skills at age 15 of
0.15 standard deviations, driven by the impacts for girls and children from
disadvantaged families. Finally, Cornelissen et al. (forthcoming) explore a policy
reform of the German child care system which entitles every child to a place on
their third birthday. They find child care attendance has a more positive effect for
children from more disadvantaged backgrounds. While there are clearly studies
7
here that report positive impacts of universal children programs, in many cases
these impacts are primarily enjoyed by less advantaged children. There is a little
clear evidence that these programs provide significant benefits more broadly.
Universal preschool has also been a focus of recent research in the United
States. Many of these studies exploit age cutoffs for preschool enrollment
comparing the youngest children in a preschool cohort to the children just a little
bit younger who had to wait an additional year before enrolling. Perhaps the best-
known program is in Oklahoma. Gormley and Gayer (2005) document positive
impacts for Hispanics and blacks, but not for whites, which is correlated with
eligibility for free school lunch. Using a different cognitive measure Gormley et
al. (2005) report more broadly-based gains. A study of New Mexico’s program
(Hustedt et al. 2008) finds positive effects on math achievement and literacy in a
sample that over represents Hispanics and Native Americans. Taking a wider
view, Wong et al. (2007) examine preschool programs in five states (a mix of
targeted and universal programs) on a variety of outcomes. They record positive
impacts on a little more than half of the outcomes investigated. Finally,
Fitzpatrick (2008) studies the introduction of pre-K program in Georgia, finding
positive impacts for disadvantaged children in small towns and rural areas. As
with the European studies, the recent American evidence mostly fits the pattern
that the positive impact of universal programs is concentrated in more at-risk
children.
Most relevant to the current paper is research on the introduction of universal
child care in Quebec. The initial evaluation of this policy in BGM found striking
negative impacts of the program on children’s non cognitive scores and family
outcomes. In a series of papers Kottelenberg and Lehrer show that most of these
contemporaneous negative effects of the program on young children and family
outcomes measured shortly after it was introduced have persisted on newer
cohorts of children as the program has matured (2013), that the negative impacts
8
on child outcomes are larger the younger the age the child entered the program
(2014), that the impacts vary by the sex of the child (2018). They also find (2017)
evidence of heterogeneous impacts. For PPVT and Motor-Social Development
scores they find more positive outcomes for children in single-parent families and
in the bottom quintiles of the tests score distribution. These are offset at the mean
by negative impacts for children from two-parent families. Haeck et al. (2015)
present evidence that the program had negative effects on children’s cognitive
development at age 5. Finally, Brodeur and Connolly (2013) and Molnar (2017)
focus on heterogeneity by education, with the former paper looking at life
satisfaction and the latter time allocation. Our work is distinguished from these
previous studies by focusing on longer-run outcomes at older ages.
To summarize, the literature on child care and preschool seems to indicate that
high-quality interventions for low-income populations deliver both short and
long-run benefits, particularly through non cognitive channels. But, universal
child care expansions do not appear to provide broadly-based short-term benefits,
with mixed evidence on long-term effects.
II. The Quebec Universal Child Care Policy
Introduced in September 1997, the Quebec child care policy aimed to provide
regulated child care places to all children aged 0-4 in the province at a price of $5
per day, with the rest of the cost covered by government subsidy. This program
raised child care subsidies to almost 80 percent on average, which can be
compared to subsidies of roughly one-third in other provinces.4 Children were
eligible for the program whether or not their parents worked. Child care under the
4 BGM report that the program raised subsidies for two-parent households from 50% to 80%;
and 60% to 80% in single-parent households. In the pre-policy period these subsidies were tax
incentives rather than direct price subsidies. A full description of child care subsidization in
Canada in these years is presented in Baker et al. (2005).
9
program was primarily provided in two venues. The first were child care centers
(centres de la petite enfance--CPE) created out of existing nonprofit child care
centers. The second was home-based care staffed by regulated providers and
organized into networks affiliated with a local CPE. Existing for-profit child care
centers could provide subsidized spaces as well. Typically, older children enrolled
in the CPE center-based care and younger children were enrolled in family home-
based care. The fee was raised to $7 a day in 2004 and to $7.30 in 2014. Haeck et
al. (2015) report that regulated child care places in the province rose from 78,864
in 1997 to 245,107 in 2012, while total provincial subsidies rose from 288 million
dollars in 1996/97 before the program to 2.2 billion dollars in 2011/12.5
The introduction of the Quebec program was accompanied by some important
structural reforms of child care provision. Formal qualifications for caregivers
were raised and operational regulations modified.6 There was also an expansion of
voluntary full-time kindergarten and subsidized after-school care for children
aged 5-12. The government introduced new higher wage policies, pressured in
part by strikes by the unionized child care providers. In our analysis, we cannot
distinguish the impacts of these supply-side interventions on the quality of care
from the subsidization of fees which happened at the same time.
The program was first introduced to four-year olds in September 1997. So,
children born before 1993 were not eligible. In 1998 three-year olds were
5 There appears to have been queues for subsidized places at the start of the program. The
magnitude of excess demand is hard to estimate, however, because waiting lists included children
not yet eligible, children in subsidized care but wanting to change providers and children on
multiple wait lists. 6 The proportion of staff required to have a college diploma or university degree in early
childhood education rose from one-third to two-thirds. To facilitate this goal the government
provided financial aid for staff enrolled in college-level early childhood education. Home-based
providers faced increased training (24–45 hours) and annual professional development (6 hours)
requirements. While maximum center size increased from 60 to 80 places, staff/child ratios
remained unchanged, except the ratios for 4- and 5-year-olds rose from 1 : 8 to 1 : 10. There was
also an increase in parental participation in governance as their representation on the board of
directors rose from 51% to two-thirds of members.
10
included, followed in 1999 by two-year olds. Finally, in 2000, both zero- and one-
year old children were included. So, birth cohorts from 2000 onward were eligible
from ages zero to four, while those born from 1993 to 1999 were eligible for part
of their early lives. Appendix Figure 1 depicts this cohort eligibility pattern.
There has been a number of reviews of the program’s quality since its
inception, which we review in the Appendix. In Japel et al. (2005), the quality of
care was judged to be at “minimal quality” in just over 60 percent of the venues
evaluated, and just over one-quarter provided services that were judged good,
very good or excellent. This is comparable to the quality of care provided in many
other developed countries, and better than the quality of care in for-profit or
unregulated care in Quebec.
III. Data and Empirical Strategy
We make use of four types of data (consisting of six data sets) for our analysis
to trace the long-run impact of the Quebec program from the period of treatment
through to young adulthood, covering a variety of relevant outcomes. For all the
data sources, our sample selection decisions are guided by how each source
covered the cohorts exposed to program treatment. Below we describe each of the
four data sources in turn.
A. Child Care Enrollment and Child Outcomes: NLSCY and SYC
Our first dataset is the National Longitudinal Study of Children and Youth
(NLSCY), which was the primary dataset in BGM. The NLSCY is a nationally
representative survey of children, conducted biannually between 1994-95 (cycle
1) and 2008-09 (cycle 8). While the NLSCY has a longitudinal component, we
11
use it cross-sectionally.7 About 2,000 children ages 0 to 5 are available in each
cycle, with some coverage of children at older ages. We also add the Survey of
Young Canadians (SYC) conducted in 2010-11 as a cross-section survey of
children aged 1-9 with similar child development content as the NLSCY. We pool
the data from the NLSCY (cycles 1 to 8) and SYC (which we denote as cycle 9)
together to produce a time series of cross sections.8
We use the NLSCY/SYC for two purposes. First, we re-examine the
contemporaneous impact of the Quebec Family Plan on child outcomes at age 5
or younger. We benchmark our estimates to BGM using cycles 1 through 5
(excluding the transitional cycle 3, as in BGM). We then extend this evidence by
adding NLSCY cycles 6-8 and the SYC. Second, we want to see if the estimated
contemporaneous impacts of the program persist into grade school. We do this by
selecting cycles that provide a cross-section view of the children at ages 5 through
9. This includes cycles 1 and 2 for the ‘pre’ period which we compare to
observations for this same age group in cycle 7 and cycle 9 (the SYC).9
At the younger ages we focus on the same outcome measures examined in
BGM and many subsequent studies of the Quebec program. We use a binary
indicator for any type of non parental care while the parent works or is at school,
a set of parent-reported non cognitive scores, the child’s score on the well-known
Peabody Picture Vocabulary Test (PPVT), and finally a parent report of how well
the child gets along at school with her/his teacher.
7 The longitudinal component started with an initial cohort of children age 0-11 in the first
cycle. These children have been followed in each subsequent cycle, with non longitudinal kids at
other ages added to fill in the gaps. 8 By focusing on restricted age intervals (e.g., ages 2-3 or 4-5) children appear only once in our
pooled samples. 9 Cycle 7 is the only post-treatment NLSCY cycle with children at each age between 5 and 9.
The other cycles (4-6 and 8) have holes at some or all ages. Our results are similar using all
available cycle 4 to cycle 8 observations within the age 5 to 9 range. Our sample restricted to cycle
7 and the SYC is a more conservative approach.
12
The non cognitive scores are built up from a menu of questions about the
behavior and development of their children. The questions are based on best
practices in the relevant fields. Each question solicits a response on a three-point
scale, and the score is constructed as the sum of responses on all questions for a
given behavior or skill. An overview of the questions that make up each measure
is provided in the Appendix.10 At ages 2 and 3 we observe scores for
Hyperactivity, Anxiety, Separation Anxiety, and Aggression.11 For the 5-9 year
olds we have scores for Hyperactivity, Anxiety, Aggression, Indirect Aggression
and Prosocial Behaviour. While some of the indices for the older age group have
the same names as corresponding indices for the younger children, they are based
on a different set of age-appropriate questions.
B. Test Scores: SAIP/PCAP and PISA
To measure the impact of the Quebec program on test scores of older children,
we turn to two different sources. The first source combines data from the School
Achievement Indicators Program (SAIP) and the successor Pan Canadian
Assessment Program (PCAP), which are initiatives of the Council of Ministers of
Education. The SAIP assesses the performance of 13 and 16 year old students
across the country in the core subjects of math, reading and science. The tests
were conducted 9 times between 1993 and 2004, each focusing on one of the core
subjects. The PCAP succeeded the SAIP and has been conducted triennially
starting in 2007; we use 2007, 2010, and 2013 tests. Each PCAP focuses on one
of the core subjects; just like SAIP. Unlike SAIP, however, a smaller sample of
students writes tests in the other non focal subjects. So, scores for each subject are
10 A more detailed discussion of the NLSCY non cognitive measures was published in the
online appendix to BGM. 11 For the non cognitive outcomes we focus on 2-3 year olds within the 0-4 age group, because
the measures do not exist for children ages 0-1 and the non cognitive indices for 4 year olds are
based on different questions.
13
available in each PCAP wave. We construct time-series cross-section samples,
pooling data from SAIP and PCAP for the test scores in each subject area. In
addition to the PCAP data, from SAIP we have math scores from 1997 and 2001,
reading scores from 1998 and science scores from 1996.
The second data set comes from the Programme for International Student
Assessment (PISA), which is a triennial test of 15-year olds conducted by the
OECD around the world. This testing program was initiated in 2000, and covers
the core subject areas of math, reading and science. Our analysis sample includes
the Canadian test scores from 2000-2015.
C. Health and Well-Being: CCHS
To assess the impact of the child care intervention on the health of older
children, we use the Canadian Community Health Survey (CCHS). The CCHS
offers biannual national cross-section data for 2001, 2003, and 2005 of
approximately 130,000 observations; followed by annual cross-section surveys of
around 65,000 observations starting in 2007. We use all surveys up to and
including 2015. We examine questions on self-assessed health, life satisfaction,
and mental health. We take a sample of 12 through 20 year olds, which in the
chosen years contains both individuals who were and were not exposed to the
child care program at younger ages.
D. Criminal Behavior: UCRS
We combine special tabulations of crime accusations and convictions from
Statistics Canada’s Uniform Crime Reporting Survey (UCRS) with single-age
population counts to construct crime rates by age, sex, province, year cells. The
14
UCRS is a survey of police-reported crime.12 This means that the crime incident
has been substantiated by the police, and therefore the survey misses crimes that
are never detected and/or not reported to the police. The accused includes those
ultimately charged as well as cases dealt with through extrajudicial measures. We
examine rates (separately) for crimes against persons, crimes against property,
“other criminal code violations”, and drug violations; as well as an aggregate
crime rate based on these four categories.13 For our age groups most “other
criminal code violations” involve failures to appear in court and breaches of
probation.14
Our data is annual for the years 2006 through 2014. We start the analysis in
2006 to avoid any impact of the introduction of the Youth Criminal Justice Act in
2003. We discuss this choice in depth below. As in our analysis of the CCHS, we
construct our sample for 12- through 20-year olds. We exclude the data from 2010
because of missing data for that year from Montreal, the largest city in Quebec.
E. Empirical Strategy
Our basic empirical strategy is a straightforward difference-in-difference
analysis that follows BGM. The base observation is of children i in province p in
year t. The analysis of a given outcome is for a specific age interval (e.g., 2-3 year
olds). We assign a binary indicator EXPOSUREpt equal to one for children who
live in Quebec, in the years that children in that age interval would be exposed to
12 Responding to the coverage is mandatory and survey compliance is reported as “virtually 100
percent” (http://www23.statcan.gc.ca/imdb/p2SV.pl?Function=getSurvey&SDDS=3302). 13 We omit the traffic crime category as it is less relevant to our age range. We also omit a
residual ‘other federal statute violations’ that includes violations of legislation such as the
Bankruptcy Act and the Competition Act. 14 Other prevalent youth crimes are theft under $5000, assault, mischief, breaking and entering,
cannabis possession, uttering threats and possession of stolen property. See Zhang (2014) for a
recent comparison of youth and adult crime rates by offence.
15
the Quebec child care program.15 So, for example, 2-3 three year olds in Quebec
in 1996-97 (wave 2 of the NLSCY) would not be exposed to the Quebec program,
while 2-3 year olds who live in Quebec in 2001-03 (cycle 5 of the NLSCY)
would. For the NLSCY/SYC and SAIP/PCAP/PISA datasets, all children in a
given year and province have the same value of EXPOSUREpt, so we estimate
basic difference-in-difference models of the form:
(1) .
We control for a set of province dummies (PROV) and year dummies (YEAR),
as well as variables X that differ (according to availability as shown in Appendix
Table 2) by data set but can include gender, child’s age, mother’s age and
education, the number of older and younger siblings, and mother’s
/father’s/family’s immigrant status and ethnicity. We focus on the estimation of ,
the coefficient on exposure to the Quebec child care program. The core
identifying assumption is that the time trend across years is common between
Quebec and other provinces. We assess this assumption graphically later in the
paper.
For the health and crime analysis, we have data that covers a larger number of
cohorts over a larger number of years. This allows us to estimate a more flexible
version of equation (1) with a more extensive set of controls. Our samples span a
larger interval of ages so that the EXPOSURE variable now varies by age (i.e.,
birth cohort) within year. The base equation estimated for the crime rate CR is
15 See the complete cohort map in Appendix Figure 1. We assume that children observed in
Quebec at older ages also lived there from age 0-4. In the 2016 Census, the proportion of Quebec
residents who were resident in another province five years previous was 1.1 percent. (Statistics
Canada Catalogue no. 98-400-X2016313.)
16
(2)
where a indexes age, s indexes sex, p indexes province, t indexes year, SEX is a
0/1 indicator for males and AGE is a vector of single-year age dummy variables.
We also estimate a variant of equation (2) in which we add a full set of second-
order interactions between PROV, AGE, SEX, and YEAR except YEAR*PROV. In
a third specification we add a set of province-specific linear trends. For samples in
which we pool the crime rates for different offences together, we add a full set of
offence fixed effects and (as noted) their interactions with AGE, PROV, SEX and
YEAR.
F. Threats to Identification
The major threat to our identification strategy comes from pre-existing time
trends that could confound our policy inference. We use three approaches to
building the case for our policy inference. First, we present graphs showing the
trends in our outcome variables in the before and after periods, comparing the rest
of Canada to Quebec. Second, we include robust controls in our regression
specifications, with province-specific time trends in our crime analysis along with
time interactions with age and sex. Finally, for the crime analysis in the appendix
we report results from regressions using leads and lags of the policy variable as a
type of placebo test to examine whether our policy variable is picking up
underlying trends.
17
G. Scaling Reduced-Form Results
As discussed in BGM, our modeling of outcomes is a reduced form of an
underlying process through which the Quebec policy impacts maternal labor
supply and child care utilization. To interpret the results structurally, BGM either
scaled the estimated effects by the impact of the Quebec policy on maternal labor
supply (a 7% rise) or by its impact on use of child care (a 14% rise). Haeck et al.
(2015) show that between the mid-1990s and 2008 the proportion of children aged
1-4 who were in center-based care as their primary arrangement rose in Quebec
from under 10 percent to close to 60 percent, while in the rest of Canada it rose
from about 10 percent to just under 20 percent. The proportion in parental care
fell from around 55 percent to roughly 25 percent in Quebec over this same
period, while the similar proportion in the rest of the country fell from just under
60 percent to about 50 percent, where it has stabilized since 1998. By this metric
the proportion of treated children in Quebec is much higher than the proportion
that moved into non parental care with the advent of the program.
There are therefore a wide variety of “first stage” estimates one could apply to
the longer-run reduced-form impacts we estimate here. As a result, we are reticent
here to interpret any of our longer run results in a structural way and focus instead
on the sign and significance of our reduced-form findings.
H. Additional Factors
In our previous study of the Quebec program we limited the analysis to children
in two-parent families. This choice minimized any possible confounding effects
of concurrent changes to Canada’s National Child Benefit on our sample of 0-4-
18
year olds.16 As we turn our focus to children at older ages, the restriction to two-
parent families makes less sense. Due to family dynamics, at older ages children
currently living in single-parent families may have lived in two-parent families
when they were young. Likewise, children currently in two-parent families may
have been born into single-parent households. We therefore sample children in all
family types.
Another factor relevant to our analysis of teenage criminal activity is that the
Youth Criminal Justice Act (YCJA) came into effect on April 1, 2003. The YCJA
is a federal act governing the prosecution of youth crimes across the country.
Quebec has a history of taking a more rehabilitative approach to youth criminal
activity. One of the impacts of the YCJA was to make the rest of Canada more
like Quebec, in that it encouraged the use of extrajudicial remedies instead of the
courts for less severe crimes.17 Correspondingly there appears to be a sharp drop
in the proportion of youth offenders charged in most provinces in 2003 and
corresponding uptick in the proportion chargeable but not charged (Carrington
and Scholenberg 2005). An exception is Quebec, no doubt reflecting the
province’s pre-existing proclivity for extrajudicial measures for youth crime. As
evidenced in Bala et al. (2009), this impact appears mostly discrete to the year the
Act was implemented, and the rates of charged and otherwise cleared youth
crimes “settled” into new post YCJA levels by about 2005. As a result we use
crime data starting in 2006 to stay clear of this impact of the YCJA. We also note
that by examining both data on accusations and convictions we provide evidence
16 BGM show the main findings extend to the children of single-parent households. Milligan
and Stabile (2011, Figure 1) show that child benefits saw about a threefold increase between 1990
and 2005 for families under CAD25,000 income, but did not increase after 1994 for families with
CAD50,000 or more of income—which is a large majority of two-parent families. Including
children from single-parent families in our sample may attenuate the estimated impacts if the child
benefits push child development in the opposite direction. 17 As argued by Trepanier (2004) the YCJA also put some limits on Quebec’s rehabilitative
approach (for example, no rehabilitation while accused is remanded in custody) and was perceived
as a triumph of the principle of proportionality over rehabilitation and reintegration.
19
of criminal activity both before and after the application of any extrajudicial
remedies.
IV. Contemporaneous Impacts
In this section we model the contemporaneous impact of exposure to the
Quebec Family Plan on child care use and non cognitive skills. We begin by
replicating earlier analysis showing a substantial increase in the use of non
parental care. We then reproduce our earlier results on the negative effects on
contemporaneous non cognitive skills of young children.
[ Insert Figure 1 Here]
We begin by graphing in Figure 1 the unconditional standardized means of our
contemporaneous dependent variables for Quebec and for the rest of Canada by
cycle of the NLSCY. We standardize the score-based dependent variables to have
mean zero and unit standard deviation. For each outcome, we indicate the onset of
the policy with a vertical line at cycle 3 in 1998-99. For the first five outcomes,
the solid line for the rest of Canada is almost flat, indicating little trend. This
stability in the untreated provinces is an important part of the case for our
identification strategy. Also, with perhaps the exception of separation anxiety, the
scores in cycles 1 and 2 in Quebec are visually parallel to the scores in the rest of
Canada. The more notable exception is the PPVT score, which shows an upward
trend starting in cycle 4 for the rest of Canada. Furthermore, for Quebec, there is a
large, anomalous downward spike in PPVT scores in cycle 3, which almost
completely dissipates in cycle 4. Post policy, being in care increases sharply
starting in cycle 3 and continuing through cycle 9. The behavioral scores also
each show a distinct relative increase after cycle 3, with varying patterns in later
cycles. The time trends for the PPVT in Quebec are less clear.
20
In Table 1 we report regression results to quantify the impacts seen in the
figure. These regressions use a similar specification to the original BGM analysis
but differ by the inclusion of both single-parent and two-parent families. We
exclude observations from cycle 3, as the program is in transition. In our
regressions, we report intent-to-treat estimates unscaled by proportion of the
population treated, for the reasons discussed above. We show standard errors both
clustered by province-year of birth and estimated following the method of Bester
et al. (2011) allowing dependence over time and within region.18 Finally the
reported levels of statistical significance are adjusted for multiple testing
following Anderson (2008).
[ Insert Table 1 Here]
In the first row are estimates of the program effect on the probability of the
child being enrolled in child care at ages zero to four. At just under 15 percentage
points, the estimate using waves 1, 2, 3 and 4 almost exactly matches the estimate
in our previous paper (0.146). Extending the sample by the additional cycles of
the NLSCY and the SYC leads to a larger estimate, which might be expected as
the supply of subsidized child care spaces expanded as the program matured. This
latter estimate represents an increase in child care use of just over 40 percent of
the baseline rate.
In the next four rows are the estimates of the program impact on standardized
non cognitive outcomes at ages 2-3. In each case we have created the variable so
that a higher score indicates a poorer outcome. The components of each score are
listed in the Appendix. The estimates using cycles 1, 2, 4, and 5 show statistically
significant estimates for Anxiety and Aggression but not for Hyperactivity and
18 For the Bester et al. (2011) method, we cluster using four regional groups and base inference
on a t -distribution with three degrees of freedom.
21
Separation Anxiety. They are marginally smaller in magnitude than the estimates
in our previous paper, although not enough to qualitatively change our
inference.19 Adding the additional waves of data leads to a statistically significant
estimate for Hyperactivity and maintains the inference for Anxiety and
Aggression. Also, the estimates are generally stable in magnitude across the two
samples, except for Aggression, which is just under 50 percent larger.
In the last row is the estimated impact of the program on a measure of cognitive
development—PPVT—at ages 4 and 5. We report a statistically-significant
decline of 11 to 14 percent of a standard deviation depending on sample. This
result, while consistent with Haeck et al. (2015), is different than the estimate in
BGM which used a different age range.20
The results in Table 1 demonstrate that the main conclusions of BGM for young
children of two-parent families extend to the full sample of young children from
all family types and persist as the program has matured. The Quebec program led
to a substantial increase in the use of child care and increases in children’s levels
of anxiety and aggression. We do not pursue analysis of heterogeneity for these
contemporaneous outcomes, but Kottelenberg and Lehrer (2017) provide
evidence of a positive boost to child development for children from disadvantaged
single-parent families, with more negative outcomes from two-parent families. In
addition, Kottelenberg and Lehrer (2018) find significant differences between
boys and girls.21
19 In BGM the estimates are 9 percent (anxiety) and 12 percent (aggression) of a standard
deviation. 20The addition of five-year olds here explains the difference to BGM. With an age 4 sample, we
obtain an insignificant estimate of -0.250 (0.843), consistent with the insignificant result of 0.36
(0.75) in BGM. 21 In the appendix we do examine the policy impact using the Firpo, Fortin, and Lemieux (2009)
unconditional quantile regression approach described below. The results indicate that the impact is
close to the mean impact across deciles for three of the four non cognitive outcomes. The notable
exception is aggression for which we find a larger impact at higher deciles—that is for children
with elevated scores.
22
V. Persistence of Impacts at Older Ages
We now turn to the examination of the persistence of the impacts at older ages.
The analysis begins with the use of several behavioral non cognitive scores at
ages 5-9, and then moves on to cognitive test scores at early teen ages. Following
that we examine health outcomes, and finally criminal behavior.
A. Non Cognitive Outcomes
In Table 2 we present the results for non cognitive outcomes at ages 5-9. The
details on each measure’s construction are provided in the Appendix. We again
normalize the non cognitive scores to have zero mean and unit standard deviation.
The first four measures show that the Quebec program’s negative effects on non
cognitive skills appear to strongly persist into school years, and in some cases are
larger than at younger ages. For Anxiety the impact is now just over one quarter
of a standard deviation, which is more than twice as large as for 2-3 year olds,
while for Aggression the estimate is very similar to the result for the younger age
group. Hyperactivity shows an increase of 13 percent of a standard deviation. For
the two new indices we see a statistically significant impact on Indirect
Aggression of 19 percent of a standard deviation, while the result for Prosocial
behavior is very close to zero.
[ Insert Table 2 Here]
For the older children we also have an alternative measure of behavior: a
parent-reported indication of how the child gets along with his/her teacher at
school. The variable is coded 0/1, where one indicates the child gets along very
well with his/her teacher (there are no problems). The estimate for this variable is
in the last row of Table 2. It is consistent with the results for the non cognitive
23
indices, indicating exposure to the Quebec program leads to a statistically
significant worse outcome.
We extend this analysis by examining heterogeneity in the policy impact using
the Firpo, Fortin, and Lemieux (2009) unconditional quantile regression approach
based on the recentered influence function (RIF). These RIF estimates show how
the policy variable affects individuals at the decile cutoffs of the unconditional
distribution of the dependent variable. We graph in Figure 2 the RIF results by
decile cutoff for the first four dependent variables analyzed in Table 2.22 All four
non cognitive scores show larger impacts at higher deciles than at lower deciles.
The impact on aggression is much larger at the 7th, 8th, and 9th decile cutoffs. This
suggests that the primary impact of the program was to increase aggression scores
for those who already had high scores.23
[ Insert Figure 2 Here]
In the appendix, we provide further analysis of these outcomes by performing
the analysis separately by gender. The results indicate that the impact on boys is
generally stronger, especially in the cases of hyperactivity and aggression.
Taken together, the negative impact of the Quebec program on the non
cognitive outcomes of young children appears to persist and in some cases
increase as they reach school ages. The impact is stronger on those who already
had elevated non cognitive scores and for boys.
22 We leave out the prosocial score since it was not statistically significant and also the binary
indicator for getting along with the teacher, since RIF analysis is only useful with continuous
dependent variables. 23
Heckman, Smith, and Clements (1997) note that impacts on quantiles need not correspond to
impacts on individuals who would otherwise be at a particular quantile.
24
B. Cognitive Outcomes
We now turn to the study of cognitive outcomes. Unfortunately, there is no
parallel verbal cognitive measure available in the NLSCY at older ages to follow
up on the PPVT result in Table 1. A math score is available, but we do not use it
here due to data issues that render it unreliable.24 Instead we use data from
periodic standardized testing of Canadian teens through SAIP/PCAP and PISA.
Note that the 2009 PISA scores for 15-year olds are likely to capture both
teenagers in Quebec who were and were not exposed to the child care program.
We consider different coding of the EXPOSURE dummy for the 2009 scores to
discover how the estimates vary on this margin.
The estimates are presented in Table 3. The standard deviations of the scores
are approximately one so the point estimates can be read directly as proportions of
a standard deviation. In the first row are the results for the PCAP/SAIP tests. The
estimates indicate negative but statistically insignificant impacts of exposure to
the Quebec program on math scores, reading, and science scores. In the next two
rows are the results for the PISA tests alternatively viewing the 2009 scores as
capturing Quebec children who are not or who are exposed to the child care
program. If we view the 2009 scores as pre-program, we obtain a statistically
significant positive impact of exposure on math scores of just over 30 percent of a
standard deviation, a marginally significant increase for reading of 10 percent and
a result for science which is statistically insignificant and close to 0. If instead we
view the 2009 scores as post-program, the impact on math is still positive but
smaller at 20 percent of a standard deviation while the impacts on reading and
science are both close to zero and statistically insignificant.
24 A disproportionate number of perfect scores (particularly in Quebec) was detected in the first
two cycles. Also, the response rate to the math test was low and variable in cycles 1-3.
25
[ Insert Table 3 Here]
On balance the results in Table 3 do not provide unambiguous evidence of a
persistent negative impact of the Quebec program on cognitive ability. For math,
the estimates show exposure to the Quebec program leading to a decrease in
scores (insignificant) in the SAIP/PCAP but an increase in PISA scores.25 The
results for reading and especially science provide a more consistent story of no
impact of the Quebec program. Overall the evidence on the long-run impact of the
Quebec Family Plan on test scores is mixed, and there is no evidence that the
initial negative impact on PPVT scores persist as evidenced by the reading scores
we examine.
C. Health and Life Satisfaction
We next study the impact of exposure to the Quebec program on health status
and on life satisfaction using the CCHS survey. The motivation is to see if the
negative impacts on physical health and behavior at early ages have a persistent
impact. We use a sample with ages 12-20 for this analysis. We examine the
outcomes self-perceived health, satisfaction with life in general, and finally self-
perceived mental health. These variables are coded between 1 and 5, running from
better to worse, and we normalized them by their standard deviations for our
regressions. As noted in the Introduction our hypotheses here are that the treated
children may enjoy higher immunity due to earlier exposure to germs, and/or the
negative shocks to their non cognitive development lead to lower levels of (self-
perceived) wellbeing.
25 The relatively stronger performance of Quebec students on PISA measured math testing is a
matter of some public debate in Canada. In Quebec a teaching certificate requires a 4 year degree
in education instead of a 1-2 year certificate as in other provinces. Also, Quebec students make the
transition to high school courses taught by subject specialists in grade 7 while in most other
provinces it is in grade 9.
26
[ Insert Table 4 Here]
The estimates in Table 4 indicate that exposure to the Quebec program is
associated with some worsening of self-reported health and overall life
satisfaction, but not in self-reported mental health. For youths exposed to the
program, the poor health indicator rises by 7.3 percent of a standard deviation.
The estimate for life satisfaction is smaller, but still statistically significant, and
again indicates a poorer outcome from exposure to the program. Finally, the
estimate for mental health is negative but very small and statistically insignificant.
This may indicate that mental health is not affected by negative shocks to non
cognitive development at earlier ages, some of the developmental impacts (e.g.,
hyperactivity) fade as the children age, or that the children are not self aware and
a clinical report would show a different result.
D. Youth Crime
Our final measure of longer-run outcomes is youth criminal activity. In
evaluations of Perry Preschool, the long-run impact on crime was a vital
component of the analysis.26 Our aim here is to investigate whether the link
between non cognitive development and crime holds up in a symmetric case
where there is a decrease in measured non cognitive development. We have two
measures of criminality—rates of accused and convictions. We focus on four
crimes (personal, property, other criminal code convictions and drugs), as well as
an aggregate measure of the incidence of all these crime categories.
26 Belfield et al. (2006) find that crime reduction by males provides most of the long-run
financial benefit of the Perry Preschool program. Heckman et al. (2013, p. 2070) find in their
study of Perry that “…the evidence from this paper suggests that reducing early externalizing
behavior reduces crime.”
27
[ Insert Figure 3 Here]
To begin, we examine time trends by year of birth cohorts with different
exposure to the policy. We graph the difference between Quebec and the rest of
Canada across ages separately by year-of-birth cohort in Figure 3. We group the
cohorts by the number of years of eligibility, ranging from 0 before 1993 birth
year to 5 for those born in 2000 or later. The differences are negative, indicating
lower accusation rates in Quebec than in the rest of Canada. The pattern across
cohorts is striking—the gap between Quebec and the rest of Canada shrinks for
the cohorts most exposed to eligibility for the Quebec program. In the appendix
we repeat the analysis for convictions and separately for four specific crime types
with similar monotonic shifts across birth cohorts consistent with the degree of
exposure.
In Table 5 we formalize this inference with regression estimates. Column (1)
presents the simple difference-in-differences results, controlling for fixed effects
for province, year, age, and gender. In addition, we include a set of dummies for
crime type in the pooled regression for all crime types. In column (2), a richer
specification includes the full set of second-order interactions between province,
age and gender (and crime type in the aggregate rate regressions). Finally, in
column (3) we add controls for province*year trends (and crime
type*province*year in the aggregate rate regressions) to allow for province-
specific trends.
[ Insert Table 5 Here]
The estimates are generally consistent with the graphical evidence: exposure to
the Quebec program leads to higher rates of crime. Looking first at all crime
counts, the estimates from the simple difference-in-differences specification
indicate increases in both the rates of accused and convictions that is statistically
28
significant. The coefficient of 514 for accused is a 27 percent increase on the
overall mean of 1872 accusations per 100,000. The coefficient of 208 for
convictions is of a similar magnitude relative to the mean conviction rate. This
estimate for rates of accused does not change much when we add the second order
province/age/gender interactions in the second column, but there is an increase in
the estimates for convictions. In column (3) the estimate for accused falls but
remains sizable and highly significant while the estimate for convictions returns
to the level seen in column (1). The estimates from the richest specifications
indicate sizeable effects on crime rates. For accused across all categories we
estimate a rise of 353 per 100,000 children, compared to a mean of 1872
accusations. This is a rise of 19 percent. The result is slightly higher in percentage
terms for convictions at 22 percent.
The remaining rows of the table show the results for each type of crime. The
impact of exposure to the Quebec program is largest for property crime; the
estimates from the richest specification show an increase in accusations of these
crimes of 602 per 100,000 children, or 19 percent of the mean, and for
convictions for these crimes of 342 per 100,000, or about 25 percent of the mean.
The estimated impacts on other criminal code violations are almost as large.
Slightly smaller are the estimated impacts on crimes against persons, at 16 percent
of the mean for both accusations and convictions. Finally, the impact for drug
crimes is 14 percent of the mean for accusations but over 23 percent of the mean
for convictions.
In the appendix, we present results separately for males and females. Gender
differences are of potential importance as there is recent evidence that the impacts
of non parental care vary by gender, as well as growing interest in gender
29
differences in childhood and adult success.27 The estimates indicate larger
absolute impacts on the crime rates for boys, particularly for other criminal code
violations and drugs.28 In fact in our richest specification some of the estimates
for girls are substantively smaller and lose some statistical significance.
Therefore, the gender differences in the impacts of the Quebec program on crime
rates line up with the gender differences in the impact of the program on non
cognitive development.
Are these findings and magnitudes credible? We consolidate in this paragraph
the case we have presented. We find large contemporaneous impacts of the
Quebec program on measures of children’s non cognitive development that persist
to school ages. These impacts are larger for boys and primarily for those who
already had elevated behavioral problems. The increase in criminal behavior we
document arises in the years and at ages that are consistent with a dose-response
relationship with eligibility for the Quebec program. Previous research has linked
improvements in similar non cognitive measures at early ages to decreased
criminal behavior; we find a symmetric effect for an early-life deterioration in
behavior. The estimated magnitude of the increase in crime in our most rigorous
specification is 19 percent of the mean for accusations and 22 percent for
convictions. For context, the overall increase in aggression at ages 5-9 was
estimated to be 17 percent of a standard deviation, but our analysis indicated this
increase is associated with children who already had elevated levels of this trait—
those who may be most at risk for future trouble. Finally, the larger increases are
for offences typical for the age—property crime and behavioral issues such as
27 See, for example, Datta Gupta and Simonsen (2010), Felfe et al. (2015), Kottelenberg and
Lehrer (2018), Baker and Milligan (2016), Bertrand and Pan (2013), Cornwell et al. (2013), Fortin
et al. (2015) and Jacob (2002). 28 Baseline crime rates are also higher for boys, but what matters here is not the share of crimes
committed by boys but whether there is more criminal activity when there is a reduction in
population non cognitive skills.
30
failure to appear in court and breaches of probation. This all said, we also
emphasize that our estimates rely on assumptions about common trends across
provinces, years, and cohorts which we have attempted to verify visually and
statistically test.
We also note that the magnitudes we report are not outside the range found in
previous studies, which also suggest large impacts of policy interventions on
crime. For example, Lochner and Moretti (2004) estimate that each year of
schooling reduces the probability of imprisonment by 0.1 percentage points (off a
base of 0.8 percent) for whites and 0.3-0.5 percentage points (off a base of 3.6
percent) for blacks. Heller et al. (2017) find that a 27-week cognitive behavioral
therapy program for teens in Chicago reduced total arrests by 28 percent and
violent crime arrests by 50 percent. More directly related to early childhood,
Heckman et al. (2013) report a drop of 2.3 arrests by age 27 for the Perry
Preschool treated group, which was a 43% drop compared to the control group.
VI. Conclusions
The rapid growth in the labor force participation of mothers of young children
has led to a strong policy interest in expanding access to non parental child care,
with a particular focus on “universal” child care availability. Although that term
has come to take many meanings, the best example in North America is clearly
the program introduced in Quebec in the late 1990s. This program made child
care much cheaper for all residents and led to an enormous expansion in use of
child care by the population. Previous work has shown this policy change led to a
large decline in measured non cognitive skills among young children exposed to
the subsidized child care. We use this initial negative shock to assess whether the
negative impact persists on longer-run outcomes in a way that mirrors findings in
31
the literature that link positive early-life interventions to positive longer-run
outcomes.
Indeed, our evidence is generally consistent with such symmetry. We find the
Quebec policy had a lasting negative impact on non cognitive skills. At older
ages, program exposure is associated with worsened health and life satisfaction,
and increased rates of criminal activity. Increases in aggression and hyperactivity
are concentrated in boys, as is the rise in the crime rates. In contrast, we find no
consistent impact on their cognitive skills.
The implications of these findings for early child care policy are potentially
profound. Our findings provide strong support for the argument that the early
childhood development environment is a crucial determinant of the long-term
success of children. This suggests that measuring the contemporaneous impact of
child care programs on development indicators is important because it is
predictive of later-life success.
For policy makers, an unanswered question is whether the evidence of negative
impacts is particular to the Quebec program, or whether the lessons here apply
more broadly. Our findings for young children clearly contrast with evaluations of
targeted programs like Perry, Abecedarian, and Head Start. The evidence we cite
suggests the quality in the Quebec program is comparable to international norms,
but likely worse quality than model programs such as Perry. Also, the Quebec
program is universal, and the more granular analysis of, for example,
Kottelenberg and Lehrer (2010) suggests that it had some positive impacts for
more disadvantaged children which are offset at the mean by negative impacts for
more advantaged families. If a universal program can be designed to improve
contemporaneous impacts on all children, our results together with evidence such
as Heckman et al. (2013) suggest such a program could lead to long-run positive
outcomes.
32
REFERENCES
Almond, Douglas, and Janet Currie. 2011. “Human Capital Development
Before Age Five.” In Handbook of Labor Economics, Volume 4 Part B, edited
by David Card and Orley Ashenfelter, 1315-1486. Amsterdam: North Holland.
Almond, Douglas, Janet Currie, and Valentina Duque. Forthcoming.
“Childhood Circumstances and Adult Outcomes: Act II.” Journal of Economic
Literature.
Almond, Douglas, Lena Edlund, and Mårten Palme. 2009. “Chernobyl's
Subclinical Legacy: Prenatal Exposure to Radioactive Fallout and School
Outcomes in Sweden.” The Quarterly Journal of Economics 124(4): 1729-
1772.
Anderson, Michael L. 2008. "Multiple Inference and Gender Differences in the
Effects of Early Intervention: A Reevaluation of the Abecedarian, Perry
Preschool, and Early Training Projects." Journal of the American Statistical
Association 103(484): 1481-1495.
Baker, Michael. 2011. “Innis Lecture: Universal Early Childhood Interventions:
What Is the Evidence Base?” Canadian Journal of Economics 44(4): 1069-105.
Baker, Michael, and Kevin Milligan. 2010. “Evidence From Maternity Leave
Expansions of the Impact of Maternal Care on Early Child Development.”
Journal of Human Resources 45(1): 1-32.
Baker, Michael, and Kevin Milligan. 2016. “Boy-Girl Differences in Parental
Time Investments: Evidence from Three Countries.” Journal of Human Capital
10(4): 399-441.
Baker, Michael, Jonathan Gruber, and Kevin Milligan. 2005. “Universal
Childcare, Maternal Labor Supply, and Family Well-Being.” National Bureau
of Economic Research Working Paper 11832.
Baker, Michael., Jonathan Gruber, and Kevin Milligan. 2008. “Universal
Child Care, Maternal Labor Supply, and Family Well-Being.” Journal of
Political Economy 116(4): 709-45.
Bala, Nicholas, Peter J. Carrington, and Julian V. Roberts. 2009. “Evaluating
the Youth Criminal Justice Act after Five Years: A Qualified Success.”
Canadian Journal of Criminology and Criminal Justice 51(2): 131-167.
Belfield, Clive R., Milagros Nores, Steve Barnett, and Lawrence
Schweinhart. 2006. “The High/Scope Perry Preschool Program: Cost-Benefit
Analysis Using Data from the Age-40 Followup.” Journal of Human Resources
41(1): 162-190.
Bertrand, Marianne, and Jessica Pan. 2013. “The Trouble with Boys: Social
Influences and the Gender Gap in Disruptive Behavior.” American Economic
Journal: Applied Economics 5(1): 32-64.
33
Bester, Alan C., Timothy G. Conley, and Christian B. Hansen. 2011.
“Inference with dependent data using cluster variance estimators.” Journal of
Econometrics 165(2): 137-151.
Bitler, Marianne P., Hilary W. Hoynes, and Thurston Domina. (2014)
“Experimental Evidence on Distributional Effects of Head Start,” NBER
Working Paper No. 20434.
Black, Sandra E., Paul J. Devereux, Katrine V. Løken, and Kjell G. Salvanes. 2014. “Care or Cash? The Effect of Child Care Subsidies on Student
Performance.” The Review of Economics and Statistics 96(5): 824-837.
Brodeur, Abel, and Marie Connolly. 2013. “Do higher child care subsidies
improve parental well-being? Evidence from Quebec’s family policies.”
Journal of Economic Behavior & Organization 93: 1–16.
Carneiro, Pedro, and James J. Heckman. 2003. “Human capital policy.” In
Inequality in America: What Role for Human Capital Policies?, edited by
James J. Heckman and Alan B. Krueger, 77-240. Cambridge, MA: MIT Press.
Carneiro, Pedro, and Rita Ginja. 2014. “Long Term Impacts of Compensatory
Pre-School on Health and Behavior: Evidence from Head Start.” American
Economic Journal: Economic Policy 6(4): 135-73.
Carrington, Peter J., and Jennifer Scholenberg. 2005. “The Impact of the
Youth Criminal Justice Act on Police Charging Practices with Young Persons:
A Preliminary Statistical Assessment.” Report to the Department of Justice
Canada, Ottawa Canada.
Cascio, Elizabeth U. 2015. “The Promises and Pitfalls of Universal Early
Education.” IZA World of Labor 116.
Cornelissen, Thomas, Christian Dustmann, Anna Raute, and Uta Schönberg. Forthcoming. “Who Benefits from Universal Child Care? Estimating Marginal
Returns to early Child Care Attendance.” Journal of Political Economy.
Cornwell, Christopher, David B. Mustard, and Jessica Van Parys. 2013.
“Noncognitive Skills and the Gender Disparities in Test Scores and Teacher
Assessments: Evidence from Primary School.” The Journal of Human
Resources 48(1): pp. 238-266.
Datta Gupta, Nabanita and Marianne Simonsen. 2010. “Non Cognitive Child
Outcomes and Universal High Quality Child Care.” Journal of Public
Economics 94(1-2): 30-43.
Felfe, Christina, Natalia Nollenberger, Núria Rodríguez-Planas. 2015. “Can’t
Buy Mommy’s Love? Universal Childcare and Children’s Long-Term
Cognitive Development.” Journal of Population Economics 28(2): 393-422.
Firpo, Sergio, Nicole M. Fortin, and Thomas Lemieux. 2009. “Unconditional
quantile regressions.” Econometrica 77(3): 953-973.
Fitzpatrick, Maria D. 2008. “Starting School at Four: The Effect of Universal
34
Prekindergarten on Children’s Academic Achievement.” B.E. Journal of
Economic Analysis and Policy 8(1): 1-40.
Fortin, Nicole M., Philip Oreopoulos, and Shelley Phipps. 2015. “Leaving
Boys Behind: Gender Disparities in High Academic Achievement.” Journal of
Human Resources 50(3): 549-579.
Gibbs, Chloe, Jens Ludwig, and Douglas L. Miller. 2013. “Head Start Origins
and Impacts.” In Legacies of the War on Poverty, edited by Martha J. Bailey
and Sheldon Danziger, 39-65. New York: Russell Sage Foundation.
Gormley, William T. Jr., and Ted Gayer. 2005. “Promoting School Readiness
in Oklahoma: An Evaluation of Tulsa’s Pre-K Program.” Journal Human
Resources 40(3): 533–58.
Gormley, William T. Jr, Ted Gayer, Deborah Phillips, and Brittany Dawson. 2005. “The Effects of Universal Pre-K on Cognitive Development,”
Developmental Psychology, 41(6): 872–84.
Haeck, Catherine, Laetitia Lebihan, and Philip Merrigan. 2018. “Universal
Child Care and Long-Term Effects on Child Well-Being: Evidence from
Canada.” Journal of Human Capital 12(1): 38-98.
Haeck, Catherine, Pierre Lefebvre, and Philip Merrigan. 2015. “Canadian
Evidence on Ten Years of Universal Preschool Policies: The Good and the
Bad,” Labour Economics 36: 137-157.
Havnes, Tarjei, and Magne Mogstad. 2011. “No Child Left Behind: Subsidized
Child Care and Children’s Long-Run Outcomes.” American Economic Journal:
Economic Policy 3(2): 97-129.
Havnes, Tarjei, and Magne Mogstad. 2014. “Is Universal Child Care Leveling
the Playing Field?” Journal of Public Economics 127: 100-114.
Heckman, James J., and Stefano Mosso. 2014. “The Economics of Human
Development and Social Mobility.” Annual Review of Economics 6(1): 689-
733.
Heckman, James J., Seong Hyeok Moon, Rodrigo Pinto, Peter A. Savelyev,
and Adam Yavitz. 2010. “The Rate of Return to the High Scope Perry
Preschool Program.” Journal of Public Economics 94: 114–28.
Heckman, James J., Rodrigo Pinto, and Peter Savelyev. 2013. “Understanding
the Mechanisms Through Which an Influential Early Childhood Program
Boosted Adult Outcomes.” American Economic Review 103(6): 2052-2086.
Heckman, James J., Jeffrey Smith, and Nancy Clements. 1997. “Making the
Most Out of Programme Evaluations and Social Experiments: Accounting for
Heterogeneity in Programme Impacts.” The Review of Economic Studies 64(4):
487-535.
Heller, Sara B., Anuj K. Shah, Jonathan Guryan, Jens Ludwig, Sendhil
35
Mullainathan, and Harold A. Pollack. 2017. “Thinking, Fast and Slow? Some
Field Experiments to Reduce Crime and Dropout in Chicago.” The Quarterly
Journal of Economics 132(1): 1-54.
Hustedt, Jason T., W. Steven Barnett, Kwanghee Jung, and Alexandra
Figueras. 2008. “Impacts of New Mexico Pre-K on Children’s School
Readiness at Kindergarten Entry: Results from the Second Year of a Growing
Initiative.” National Institute for Early Education Research, Rutgers University.
Jacob, Brian A. 2002. “Where the Boys Aren’t: Non Cognitive Skills, Returns to
School and the Gender Gap in Higher Education.” Economics of Education
Review 21(6): 589–598.
Japel, Christa, Richard E. Tremblay, and Sylvana Cote. 2005. “Quality
Counts: Assessing the Quality of Daycare Services Based on the Quebec
Longitudinal Study of Child Development” IRPP Choices 11(5): 1-42.
Kline, Patrick, and Christopher R. Walters. 2016. “Evaluating Public
Programs with Close Substitutes: The Case of Head Start,” The Quarterly
Journal of Economics 131(4): 1795-1848.
Kottelenberg, Michael J., and Steven F. Lehrer. 2013. “New Evidence on the
Impacts of Access to and Attending Universal Child-Care in Canada.”
Canadian Public Policy 39(2): 263-285.
Kottelenberg, Michael J., and Steven F. Lehrer. 2014. “Do the Perils of
Universal Child Care Depend on the Child’s Age?” CESifo Economic Studies
60(2): 338-365.
Kottelenberg, Michael J., and Steven F. Lehrer. 2017. "Targeted or Universal
Coverage? Assessing Heterogeneity in the Effects of Universal Child Care."
Journal of Labor Economics 35(3): 609-653.
Kottelenberg, Michael J., and Steven F. Lehrer. 2018. “Does Quebec's
Subsidized Child Care Policy Give Boys and Girls an Equal Start?” Canadian
Journal of Economics 51(2): 627-659.
Lefebvre, Pierre, and Philip Merrigan. 2008. “Childcare Policy and the Labor
Supply of Mothers with Young Children: A Natural Experiment from Canada.”
Journal of Labor Economics 26(3): 519–548.
Lefebvre, Pierre, Philip Merrigan, and Matthieu Verstraete. 2009. "Dynamic
Labour Supply Effects of Childcare Subsidies: Evidence from a Canadian
Natural Experiment on Universal Child Care." Labour Economics 16(5): 490-
502.
Lochner, Lance. 2011. “Nonproduction Benefits of Education: Crime, Health,
and Good Citizenship.” In Handbook of the Economics of Education, Vol. 4,
edited by Eric A. Hanushek, Stephen Machin, and Ludger Woessmann, 183-
282. Amsterdam: Elsevier Science.
Lochner, Lance, and Enrico Moretti. 2004. “The Effect of Education on Crime:
36
Evidence from Prison Inmates, Arrests, and Self-Reports.” American Economic
Review 94(1): 155-189.
Milligan, Kevin, and Mark Stabile. 2011. "Do Child Tax Benefits Affect the
Wellbeing of Children? Evidence from Canadian Child Benefit Expansions.”
American Economic Journal: Economic Policy 3(3): 175-205.
Molnar, Timea. 2017. “How Do Mothers Manage? Universal Daycare, Child
Skill Formation, and the Parental Time-Education Puzzle.” Ph.D. Thesis,
University of British Columbia.
Trepanier, Jean. 2004. “What Did Quebec Not Want: Opposition to the
Adoption of the Youth Crime Justice Act in Quebec.” Canadian Journal of
Criminology and Criminal Justice 46(3): 273-299.
Vandell Deborah Lowe, Jay Belsky, Margaret Burchinal, Nathan
Vandergrift, and Lawrence Steinberg. 2010. “Do Effects of Early Child Care
Extend to Age 15 Years? Results from the NICHD Study of Early Child Care
and Youth Development.” Child Development, 81(3): 737-756.
Wong, Vivian C., Thomas D. Cook, W. Steven Barnett, and Kwanghee Jung. 2007. An Effectiveness-based Evaluation of Five State Pre-Kindergarten
Programs using Regression-Discontinuity. New Brunswick, NJ: National
Institute for Early Education Research.
Zhang, Haimin. 2014. “Immigration and Crime: Evidence from Canada,”
Canadian Labour Market and Skills Research Network Working Paper 135.
37
FIGURE 1. TIME TRENDS IN STANDARDIZED PRESCHOOL OUTCOMES
Notes: Authors’ calculations from NLSCY/SYC data. The graph shows the mean standardized value for each of the
outcomes in Table 1 across time in Quebec and the rest of Canada. The variable in care shows the deviation from the
average value of the variable measured across all children age 0-4. The other variables are standardized using the mean and
standard deviation.
38
FIGURE 2. RIF REGRESSION ANALYSIS OF NON COGNITIVE SCORES, AGES 5-9
Notes: Authors’ calculations from NLSCY (cycles 1, 2, 7) and the SYC (cycle 9). Displayed are the mean and RIF
estimates for longer-run outcomes at ages 5-9. The dependent variable is scaled to mean zero and a unit standard deviation,
so the estimates can be interpreted as fractions of a standard deviation. The mean estimates are from Table 2. The RIF
estimates vary by decile cutoff and are described in the text. We show shaded 95 percent confidence intervals around each
estimate.
39
FIGURE 3. QUEBEC-REST OF CANADA DIFFERENCES IN ACCUSATION RATES BY BIRTH COHORT
Notes: Authors’ calculations from UCR data. Displayed is the difference in the annual accusation rate for all crime types
per 100,000 of population between Quebec and the Rest of Canada by age. Each line shows a different set of birth cohorts,
arranged by years of exposure to eligibility.
40
TABLE 1—IMPACT OF EXPOSURE TO THE QUEBEC FAMILY PLAN ON CHILD CARE AND MEASURES OF NON
COGNITIVE AND COGNITIVE OUTCOMES AT YOUNG AGES
Outcome Mean EXPOSURE
Cycles 1,2,4,5 1,2,4,5 1,2,4-9
In Care 0.46 (0.50)
0.143 [0.031]***
(0.008)***
0.197 [0.026]***
(0.011)***
Hyperactivity
2.86
(2.12)
0.065
[0.048]
(0.016)**
0.070
[0.035]*
(0.023)**
Anxiety 1.25
(1.50)
0.115
[0.027]*** (0.013)***
0.119
[0.033]*** (0.013)***
Separation Anxiety 2.76 (2.03)
0.073 [0.047]
(0.021)***
0.063 [0.042]
(0.016)**
Aggression 4.98
(2.95)
0.117
[0.040]**
(0.024)**
0.172
[0.043]***
(0.006)***
PPVT 99.98
(15.29)
-0.109
[0.045]** (0.039)*
-0.139
[0.041]*** (0.026)***
Notes: Authors’ calculations from NLSCY/SYC data. Sample includes all families. The sample ages are 0-4 years for In
Care; 2-3 years for (standardized) Hyperactivity, Anxiety, Separation Anxiety and Aggression; and ages 4-5 for
(standardized) PPVT. Reported is the coefficient on a dummy indicating exposure to eligibility. Robust standard errors
clustered on province and year of birth are in square brackets. Standard errors using the Bester et al. (2011) method are
reported in round brackets. Significance is reported using p-values adjusted for multiple testing.
*** Significant at the 1 percent level.
** Significant at the 5 percent level.
* Significant at the 10 percent level.
41
TABLE 2—IMPACT OF EXPOSURE TO THE QUEBEC FAMILY PLAN ON MEASURES OF NON COGNITIVE OUTCOMES, AGES 5-9
Outcome Mean EXPOSURE
Hyperactivity 4.09
(3.14)
0.130
[0.050]** (0.008)***
Anxiety 2.49 (2.30)
0.278 [0.041]***
(0.022)***
Aggression 1.36
(1.82)
0.169
[0.039]***
(0.009)***
Indirect Aggression 1.00
(1.55)
0.190
[0.017]*** (0.021)***
Prosocial 13.48 (3.90)
0.006 [0.048]
(0.038)
Child gets along with
Teacher (parent report)
0.80
(0.40)
-0.084
[0.023]***
(0.009)***
Hyperactivity 4.09
(3.14)
0.130
[0.050]** (0.008)***
Notes: Authors’ calculations from NLSCY (cycles 1, 2, 7) and the SYC (cycle 9). Sample includes all families. Reported is
the coefficient on a dummy indicating exposure to eligibility. Robust standard errors clustered on province and year of birth are in square brackets. Standard errors using the Bester et al. (2011) method are reported in round brackets.
Significance is reported using p-values adjusted for multiple testing.
*** Significant at the 1 percent level. ** Significant at the 5 percent level.
* Significant at the 10 percent level.
42
TABLE 3—IMPACT OF EXPOSURE TO THE QUEBEC FAMILY PLAN ON STANDARDIZED TEST SCORES
Math Reading Science
Mean EXPOSURE Mean EXPOSURE Mean EXPOSURE
SAIP/PCAP
0.125
(0.986)
-0.181
[0.118]
(0.131)
0.107
(1.000)
-0.237
[0.181]
(0.090)
0.060
(0.990)
-0.062
[0.059]
(0.044)
PISA
(2009 control)
0.119
(0.998)
0.312
[0.051]***
(0.015)***
0.144
(0.973)
0.106
[0.051]*
(0.018)**
0.122
(0.991)
0.043
[0.083]
(0.014)*
PISA
(2009 treated)
0.119
(0.998)
0.199
[0.084]*
(0.010)***
0.144
(0.973)
0.054
[0.048]
(0.012)**
0.122
(0.991)
0.021
[0.086]
(0.021)
Notes: Authors’ calculations from SAIP/PCAP and PISA test score data. Sample includes all families. Reported is the coefficient on a dummy indicating exposure to eligibility. Robust standard errors clustered on province and year of birth
are in square brackets. Robust standard errors clustered on province and year of birth are in square brackets. Standard
errors using the Bester et al. (2011) method are reported in round brackets. Significance is reported using p-values adjusted for multiple testing.
*** Significant at the 1 percent level.
** Significant at the 5 percent level. * Significant at the 10 percent level.
43
TABLE 4—IMPACT OF EXPOSURE TO THE QUEBEC FAMILY PLAN ON SELF-REPORTED HEALTH OUTCOMES
Age 12-20
Mean EXPOSURE
Health 2.09
(0.85)
0.073
[0.019]***
(0.005)***
Life Satisfaction 1.61
(0.62)
0.043
[0.017]** (0.006)***
Mental Health 1.90 (0.88)
-0.011 [0.020]
(0.005)*
Notes: Authors’ calculations from CCHS data. Sample includes all families. Reported is the coefficient on a dummy
indicating exposure to eligibility. Robust standard errors clustered on province and year of birth are in square brackets. Robust standard errors clustered on province and year of birth are in square brackets. Standard errors using the Bester et al.
(2011) method are reported in round brackets. Significance is reported using p-values adjusted for multiple testing.
*** Significant at the 1 percent level. ** Significant at the 5 percent level.
* Significant at the 10 percent level.
44
TABLE 5—IMPACT OF EXPOSURE TO THE QUEBEC FAMILY PLAN ON CRIME RATES, AGES 12-20
Mean (1) (2) (3)
Panel A. Accused
All
1872
514 [77]***
(181)**
590 [60]***
(71)**
353 [69]***
(35)***
Person
1839
530
[84]***
(171)*
649
[68]***
(90)**
299
[81]***
(70)**
Property
3100
566
[100]*** (190)*
932
[132]*** (193)**
602
[172]*** (11)***
Other CC
1604
639 [121]***
(396)
563 [64]***
(79)**
379 [68]***
(110)*
Drugs
945
322
[73]***
(54)**
217
[27]***
(67)**
130
[28]***
(41)*
Panel B. Convictions
All
963
208 [37]***
(47)**
323 [39]***
(26)***
212 [44]***
(5)***
Person
1002
291
[51]***
(62)**
323
[43]***
(11)***
167
[59]***
(27)** Property
1363
112
[61]*
(30)**
455
[70]***
(41)***
342
[93]***
(31)***
Other CC
1059
289
[57]*** (117)*
311
[50]*** (100)*
239
[54]*** (15)***
Drugs 429 140 [27]***
(35)**
203 [26]***
(43)**
99 [29]***
(25)**
Notes: Authors’ calculations from UCR data. Sample includes all families. In rows titled (1) are estimates from the
difference in differences specification. In rows titled (2) are estimates that add all second order province, age, gender, year
interactions, expect year*prov. In rows titled (3) are estimates that add province, year linear trend interactions. Reported is the coefficient on a dummy indicating exposure to eligibility. Robust standard errors clustered on province and year of
birth are in square brackets. Standard errors using the Bester et al. (2011) method are reported in round brackets.
Significance is reported using p-values adjusted for multiple testing.
*** Significant at the 1 percent level.
** Significant at the 5 percent level.
* Significant at the 10 percent level.